You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

100 lines
4.2KB

  1. #include "../dsputil.h"
  2. #include "../mpegvideo.h"
  3. #include "../avcodec.h"
  4. static void dct_unquantize_h263_intra_iwmmxt(MpegEncContext *s,
  5. DCTELEM *block, int n, int qscale)
  6. {
  7. int level, qmul, qadd;
  8. int nCoeffs;
  9. DCTELEM *block_orig = block;
  10. assert(s->block_last_index[n]>=0);
  11. qmul = qscale << 1;
  12. if (!s->h263_aic) {
  13. if (n < 4)
  14. level = block[0] * s->y_dc_scale;
  15. else
  16. level = block[0] * s->c_dc_scale;
  17. qadd = (qscale - 1) | 1;
  18. }else{
  19. qadd = 0;
  20. level = block[0];
  21. }
  22. if(s->ac_pred)
  23. nCoeffs=63;
  24. else
  25. nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
  26. __asm__ __volatile__ (
  27. /* "movd %1, %%mm6 \n\t" //qmul */
  28. /* "packssdw %%mm6, %%mm6 \n\t" */
  29. /* "packssdw %%mm6, %%mm6 \n\t" */
  30. "tbcsth wr6, %[qmul] \n\t"
  31. /* "movd %2, %%mm5 \n\t" //qadd */
  32. /* "packssdw %%mm5, %%mm5 \n\t" */
  33. /* "packssdw %%mm5, %%mm5 \n\t" */
  34. "tbcsth wr5, %[qadd] \n\t"
  35. "wzero wr7 \n\t" /* "pxor %%mm7, %%mm7 \n\t" */
  36. "wzero wr4 \n\t" /* "pxor %%mm4, %%mm4 \n\t" */
  37. "wsubh wr7, wr5, wr7 \n\t" /* "psubw %%mm5, %%mm7 \n\t" */
  38. "1: \n\t"
  39. "wldrd wr2, [%[block]] \n\t" /* "movq (%0, %3), %%mm0 \n\t" */
  40. "wldrd wr3, [%[block], #8] \n\t" /* "movq 8(%0, %3), %%mm1 \n\t" */
  41. "wmulsl wr0, wr6, wr2 \n\t" /* "pmullw %%mm6, %%mm0 \n\t" */
  42. "wmulsl wr1, wr6, wr3 \n\t" /* "pmullw %%mm6, %%mm1 \n\t" */
  43. /* "movq (%0, %3), %%mm2 \n\t" */
  44. /* "movq 8(%0, %3), %%mm3 \n\t" */
  45. "wcmpgtsh wr2, wr4, wr2 \n\t" /* "pcmpgtw %%mm4, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 */
  46. "wcmpgtsh wr3, wr4, wr2 \n\t" /* "pcmpgtw %%mm4, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 */
  47. "wxor wr0, wr2, wr0 \n\t" /* "pxor %%mm2, %%mm0 \n\t" */
  48. "wxor wr1, wr3, wr1 \n\t" /* "pxor %%mm3, %%mm1 \n\t" */
  49. "waddh wr0, wr7, wr0 \n\t" /* "paddw %%mm7, %%mm0 \n\t" */
  50. "waddh wr1, wr7, wr1 \n\t" /* "paddw %%mm7, %%mm1 \n\t" */
  51. "wxor wr2, wr0, wr2 \n\t" /* "pxor %%mm0, %%mm2 \n\t" */
  52. "wxor wr3, wr1, wr3 \n\t" /* "pxor %%mm1, %%mm3 \n\t" */
  53. "wcmpeqh wr0, wr7, wr0 \n\t" /* "pcmpeqw %%mm7, %%mm0 \n\t" // block[i] == 0 ? -1 : 0 */
  54. "wcmpeqh wr1, wr7, wr1 \n\t" /* "pcmpeqw %%mm7, %%mm1 \n\t" // block[i] == 0 ? -1 : 0 */
  55. "wandn wr0, wr2, wr0 \n\t" /* "pandn %%mm2, %%mm0 \n\t" */
  56. "wandn wr1, wr3, wr1 \n\t" /* "pandn %%mm3, %%mm1 \n\t" */
  57. "wstrd wr0, [%[block]] \n\t" /* "movq %%mm0, (%0, %3) \n\t" */
  58. "wstrd wr1, [%[block], #8] \n\t" /* "movq %%mm1, 8(%0, %3) \n\t" */
  59. "add %[block], %[block], #16 \n\t" /* "addl $16, %3 \n\t" */
  60. "subs %[i], %[i], #1 \n\t"
  61. "bne 1b \n\t" /* "jng 1b \n\t" */
  62. :[block]"+r"(block)
  63. :[i]"r"((nCoeffs + 8) / 8), [qmul]"r"(qmul), [qadd]"r"(qadd)
  64. :"memory");
  65. block_orig[0] = level;
  66. }
  67. #if 0
  68. static void dct_unquantize_h263_inter_iwmmxt(MpegEncContext *s,
  69. DCTELEM *block, int n, int qscale)
  70. {
  71. int nCoeffs;
  72. assert(s->block_last_index[n]>=0);
  73. if(s->ac_pred)
  74. nCoeffs=63;
  75. else
  76. nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
  77. ippiQuantInvInter_Compact_H263_16s_I(block, nCoeffs+1, qscale);
  78. }
  79. #endif
  80. void MPV_common_init_iwmmxt(MpegEncContext *s)
  81. {
  82. if (!(mm_flags & MM_IWMMXT)) return;
  83. s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_iwmmxt;
  84. #if 0
  85. s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_iwmmxt;
  86. #endif
  87. }